Improve: Token-In Token-Out Usage for RLHF (#2843)

This commit is contained in:
Shi Shuai
2025-01-11 23:14:26 +00:00
committed by GitHub
parent 197cbf9bab
commit c4f9707e16
12 changed files with 166 additions and 128 deletions

View File

@@ -323,9 +323,7 @@ class BatchTokenIDOut:
decoded_texts: List[str]
decode_ids: List[int]
read_offsets: List[int]
# Only used when --return-token-ids` is set
origin_input_ids: Optional[List[int]]
# Only used when `--skip-tokenizer-init` or `--return-token-ids` is set
# Only used when `--skip-tokenizer-init` is on
output_ids: Optional[List[int]]
# Detokenization configs
skip_special_tokens: List[bool]
@@ -356,10 +354,6 @@ class BatchStrOut:
# The output decoded strings
output_strs: List[str]
# The token ids
origin_input_ids: Optional[List[int]]
output_ids: Optional[List[int]]
# Token counts
# real input and output tokens can be get from
# origin_input_ids and output_ids by enabling --return_token_ids